peterc3dev

The short version: AMD shipped a chip with three processors (CPU, GPU, NPU) sharing one memory bus, and every ML runtime only uses one at a time. This project coordinates all three with learned routing — it profiles each operation, assigns it to the fastest device, and adapts over time.

Key findings: - Vulkan outperforms ROCm by ~60% on the Radeon 890M (hipMallocManaged is broken on gfx1150, Vulkan sees full memory pool) - The NPU kernel driver had an init-order bug — we patched it to get Llama 3.2 1B running at 40-46 tok/s at 2W - The personality database correctly learns routing after ~5 runs (embed→NPU, matmul→GPU, tokenize→CPU) - PyTorch had to be built from source — no pre-built wheels exist for gfx1150 (RDNA 3.5)